PhaseFormer: From Patches to Phases for Efficient and Effective Time Series Forecasting

Official implementation of PhaseFormer, a novel time series forecasting model that shifts the paradigm from traditional temporal patches to the phase domain.

This repository has been refactored for full compatibility with the Time-Series Library framework, ensuring seamless integration for researchers and practitioners. For the original implementation, please visit neumyor/PhaseFormer.

---

🚀 Key Features

Unlike standard segment-based forecasting, PhaseFormer leverages the periodic nature of time series data through:

• Phase Tokenization: Segments time series into fixed cycles based on period_len, encoding each cycle as a discrete phase token.
• Cross-Phase Interaction: Utilizes a specialized routing mechanism to facilitate information exchange between different phases.
• Phase Block Stacking: Employs deep layers to predict future phase states, which are then reconstructed back into the temporal domain.

The result? Significant improvements in both computational efficiency and forecasting accuracy.

---

🛠️ Usage

Environment Setup

Ensure you have the dependencies of the Time-Series Library installed. Then, you can run the provided scripts directly.

Running Examples

Example script are provided in the /scripts directory. You can run PhaseFormer on Traffic by:

bash ./scripts/run_traffic.py

More scripts can be found in neumyor/PhaseFormer.

---

📚 Cite This Work

If you find PhaseFormer useful in your research, please cite our ICLR 2026 paper:

@inproceedings{niu2026phaseformer,
  title={PhaseFormer: From Patches to Phases for Efficient and Effective Time Series Forecasting},
  author={Niu, Yiming and Deng, Jinliang and Tong, Yongxin},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}